The Autonomy Slider

In the [YOU] on AI Field Guide

The autonomy slider is the cycle’s most practical answer to the question [YOU] on AI keeps circling: how does the human stay at the center when the machine is doing more and more of the work? The slider’s calibration principle—match autonomy to the cost of verification—operationalizes the orange pill’s insistence that the human’s role is not to compete with the machine but to direct it, verify it, and hold the judgment the machine cannot supply.

The Trivandrum training session Edo Segal describes is the autonomy slider moved thoughtfully upward for a specific team in a specific domain: the engineers moved from full ownership of every line of code (slider at zero) to supervision of architecturally significant decisions while delegating mechanical implementation (slider in the middle). The week of working alongside them was precisely the calibration work the slider demands: establishing where the system could be trusted and where human judgment remained essential.

The concept also reframes vibe coding, Karpathy’s 2025 term for the experience of building software by talking to an AI assistant and barely looking at the code being produced. Vibe coding is the slider pushed all the way to the right for a specific context—a throwaway project, a weekend experiment, a prototype where the stakes are low and the cost of a bug is an annoyance rather than a catastrophe. It is not how production-grade, safety-critical software should be built. The viral spread of the phrase flattened this nuance; the autonomy slider restores it. Wisdom consists in knowing which slider position a given context requires.

Origin

Karpathy introduced the autonomy slider concept to recover the nuance that his 2025 vibe coding coinage had lost in circulation. The term vibe coding had been taken up as an uncomplicated celebration of AI-assisted development—a justification for accepting AI-generated code without review in all contexts. Karpathy insisted on the distinction between play and production, between exploration and engineering, and the autonomy slider was his instrument for making that distinction precise.

The concept extends Karpathy’s long-standing argument that verification is the central bottleneck of AI deployment. The reason you cannot simply hand everything to the machine is that you cannot yet trust the machine, and the reason you cannot fully trust it is that verifying its output is often as hard as producing it. The slider, properly understood, is calibrated by the cost of checking: where verification is cheap, the slider can be high; where verification is expensive and mistakes are hard to detect and harmful when missed, the slider must be low. This is engineering wisdom dressed in a memorable phrase.

Key Ideas

Verification is the calibration principle. The correct position on the autonomy slider is not determined by the general capability of the AI system or by abstract trust in the technology. It is determined by how hard it is to verify that the machine did the right thing in the specific domain, and what happens if it did not. A system that writes a flawed paragraph can be asked to try again. A system that deletes the wrong files, sends the wrong message, or spends money it should not have spent cannot be reversed as easily. The cost of an undetected mistake rises precisely as we hand the system more autonomy, which is why the slider position must be set conservatively in any domain where verification is expensive or errors are consequential.

The centaur model. At the slider’s productive middle positions, the image is of a centaur: human judgment fused with machine speed and scale, each contributing what it does best. The human provides direction, goal-setting, and verification; the machine provides tireless execution and breadth. This is more demanding than full automation, because it requires designing the interfaces and workflows that let human and machine collaborate effectively. But it is more honest about where the technology actually is, and more aligned with the structural fact that human judgment becomes more valuable, not less, as the cost of execution falls.

The decade of agents. Karpathy calls the coming period the decade of agents—choosing a decade deliberately as a rebuke to those announcing the age of autonomous AI agents has already arrived. The demonstrations are impressive; the gap between demonstration and reliable deployment in consequential settings is the same gap he learned to respect at Tesla Autopilot. An agent acting in the world inherits all the jaggedness of the underlying model and adds the higher stakes of action. Progress during the decade of agents will be measured in how much autonomy can be responsibly extended, domain by domain, as the systems earn trust—which is to say, progress will be measured by how thoughtfully humans manage the slider.

Partial autonomy as the target state. The path forward is not a leap to fully autonomous systems that operate without oversight but a gradual extension of the autonomy dial as systems earn trust in domain after domain, with humans supervising and verifying along the way. Karpathy calls this partial autonomy, and he treats it not as a transitional arrangement but as the appropriate permanent relationship between humans and AI systems in any domain where the cost of error is significant. The fantasy of fully autonomous agents that need no oversight is, in his view, premature given how unreliable the systems remain and how expensive verification becomes in exactly the domains where full autonomy would be most appealing.